Anyone Can, Almost Nobody Does

Agent capability is widening with every model generation; agent accessibility is not. For most people the entire agentic repertoire of contemporary AI remains reachable through a single surface, the chat box, while everything beyond it (standing monitors, scheduled work, conditional pipelines) requires developer-grade assembly. This paper names the inferential move that excuses the gap and proposes the design primitive that would close it. The affordance fallacy is the slide from 'X is possible for a sufficiently capable user' to 'X is accessible': a de-relativisation of Gibson's ability-relative affordance, and a domain-specific instance of Binary Bias operating on the can-do axis, by which a continuous distribution of realised uptake is collapsed to a single bit of theoretical possibility, the bit being set by the most capable observer in the room. The worked specimen is the Pi-hole: free, documented, genuinely beneficial, and installed by almost nobody. Against the gap the paper proposes the daemon, an atomic agent-feature shaped as trigger, condition, action: small enough to state in a sentence, which is precisely what makes it the unit a humane interface can wrap. Two design commitments make the primitive trustworthy and epistemically clean: the reversibility line (daemons detect, prepare, and stage unattended, but gate every irreversible, public, or under-the-user's-name action behind a one-tap confirmation) and turn-close firing (a daemon classifies committed text, never live probability clouds). The closing argument raises the stakes: under the dannelse-as-X-risk lens, accessibility is not a UX nicety but a distribution variable for population-scale metacognitive capacity, and the affordance fallacy is the mechanism by which its regressive distribution gets excused.